Thyroid nodule classification in ultrasound uses TI-RADS — a standardized system where radiologists score visual features (composition, echogenicity, shape, margin, echogenic foci) to estimate malignancy risk. Machine learning can classify nodules from images, but the features it learns are statistical patterns, not clinical descriptors. The model's “reason” for a classification doesn't map onto the vocabulary radiologists use.
The authors (arXiv:2603.22692) build a radiomics dictionary anchored to TI-RADS categories. Each extracted feature is explicitly mapped to the clinical scoring system — this texture feature corresponds to echogenicity, that shape descriptor corresponds to margin irregularity. The classification uses interpretable features that a radiologist can read, verify, and disagree with.
The through-claim: explainability in medical AI isn't about explaining the model — it's about speaking the clinician's language. A feature importance plot showing “texture_contrast” has 0.3 weight is technically transparent but clinically opaque. Mapping features to TI-RADS categories makes the model's reasoning auditable in the framework the clinician already uses. The interpretability is in the dictionary, not the algorithm.